在编写优化的ftol
函数时,我发现GCC 4.6.1
存在一些非常奇怪的行为。首先让我展示一下代码(为了清晰起见,我标记了不同之处):
fast_trunc_one, C:
int fast_trunc_one(int i) {
int mantissa, exponent, sign, r;
mantissa = (i & 0x07fffff) | 0x800000;
exponent = 150 - ((i >> 23) & 0xff);
sign = i & 0x80000000;
if (exponent < 0) {
r = mantissa << -exponent; /* diff */
} else {
r = mantissa >> exponent; /* diff */
}
return (r ^ -sign) + sign; /* diff */
}
fast_trunc_two,C:
int fast_trunc_two(int i) {
int mantissa, exponent, sign, r;
mantissa = (i & 0x07fffff) | 0x800000;
exponent = 150 - ((i >> 23) & 0xff);
sign = i & 0x80000000;
if (exponent < 0) {
r = (mantissa << -exponent) ^ -sign; /* diff */
} else {
r = (mantissa >> exponent) ^ -sign; /* diff */
}
return r + sign; /* diff */
}
看起来一样?GCC不这么认为。使用gcc -O3 -S -Wall -o test.s test.c
编译后,汇编输出如下:
fast_trunc_one,生成的:
_fast_trunc_one:
LFB0:
.cfi_startproc
movl 4(%esp), %eax
movl $150, %ecx
movl %eax, %edx
andl $8388607, %edx
sarl $23, %eax
orl $8388608, %edx
andl $255, %eax
subl %eax, %ecx
movl %edx, %eax
sarl %cl, %eax
testl %ecx, %ecx
js L5
rep
ret
.p2align 4,,7
L5:
negl %ecx
movl %edx, %eax
sall %cl, %eax
ret
.cfi_endproc
fast_trunc_two,生成的:
_fast_trunc_two:
LFB1:
.cfi_startproc
pushl %ebx
.cfi_def_cfa_offset 8
.cfi_offset 3, -8
movl 8(%esp), %eax
movl $150, %ecx
movl %eax, %ebx
movl %eax, %edx
sarl $23, %ebx
andl $8388607, %edx
andl $255, %ebx
orl $8388608, %edx
andl $-2147483648, %eax
subl %ebx, %ecx
js L9
sarl %cl, %edx
movl %eax, %ecx
negl %ecx
xorl %ecx, %edx
addl %edx, %eax
popl %ebx
.cfi_remember_state
.cfi_def_cfa_offset 4
.cfi_restore 3
ret
.p2align 4,,7
L9:
.cfi_restore_state
negl %ecx
sall %cl, %edx
movl %eax, %ecx
negl %ecx
xorl %ecx, %edx
addl %edx, %eax
popl %ebx
.cfi_restore 3
.cfi_def_cfa_offset 4
ret
.cfi_endproc
这是一个极端的差异。这实际上也在个人资料中显示出来,fast_trunc_one
比fast_trunc_two
快约30%。现在我的问题是:是什么导致了这种情况?
-S -O3 -da -fdump-tree-all
进行编译。这将创建许多中间表示的快照。依次查看它们(它们是按编号排列的),你应该能够在第一个案例中找到缺失的优化。 - zwolint
改为unsigned int
,看看差异是否消失。 - zwol(r + shifted) ^ sign
与r + (shifted ^ sign)
并不相同。我猜这使得优化器感到困惑?值得一提的是,MSVC 2010(16.00.40219.01)生成的清单几乎完全相同:https://gist.github.com/2430454 - DCoder